Skip to content

User Test Report: b11c768 Transformer Model vs b28c512nbt on RTX 3080 #1224

Description

@GuangHui68

Test Environment

Item | Details

Models tested | b11c768h12nbt3tflrs-fson-silu vs kata1-zhizi-b28c512nbt

GUI | Sabaki 0.60

Hardware | NVIDIA RTX 3080, CUDA 12.8, TensorRT 10.9

Time control | 20 seconds per move per side, pondering disabled, faster play in simple positions

Max visits setting | b11c768: 30,000; b28c512nbt: 45

b11c768 vs b28c512.zip

,000


Game Observation

Under 6.5 komi (Chinese rules), black-playing b11c768 maintained a slight lead from the very beginning. The score difference did not widen significantly for most of the game — it remained relatively close for a long period. Still, a clear trend was observed: the gap was gradually but steadily expanding in favor of b11c768.

It wasn't until around move 150 that the score lead finally stretched to about 3 points. After black's move 159, the white player (b28c512) judged the position as hopeless and resigned immediately. This is consistent with an earlier test where b11c768 also won by a large margin after a tactical breakdown by the opponent.


Key Observations

  • b11c768 showed consistent positional pressure throughout the mid-game, demonstrating superior reading depth and strategic judgment.

  • Despite the close score for most of the game, the new model's strength became increasingly apparent as the game progressed.

  • The resignation at move 159 reflects b11c768's decisive killing power in late middle-game situations.


Overall Conclusion

The new b11c768 model is remarkably strong — both in terms of raw playing strength and computational efficiency. Its significantly smaller size (approx. 70M params, vs b28c512 at approx. 280M params) makes it great news for all GPU users, as it delivers top-tier performance with far lower hardware demands.


Additional Observation: Search Behavior

During post-game analysis, one notable behavioral difference was observed:

  • The b11c768 model tends to distribute its search across a wider range of candidate moves, allocating computation to many points on the board — even those that are not obviously urgent.

  • In contrast, older models like b28c512 are more focused, concentrating their search on a smaller set of top candidates.

This suggests that the new Transformer-based model explores more broadly during analysis, which may contribute to its strength in complex, global positions, but also means its search behavior is qualitatively different from the older CNN-based models.


Final Recommendation

Highly recommended. The b11c768 model offers:

  • Faster startup time

  • Efficient computation per visit

  • Strong mid-game and late-game reading

  • Much smaller model size, making it accessible to a wider range of hardware

For anyone currently using older CNN-based models (b28, b40, etc.), upgrading to v1.17.0+ with the b11c768 Transformer model is a clear and worthwhile step forward.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions